The Monetary Authority of Singapore is testing whether cross-institutional transaction data, processed through artificial intelligence models, can identify scam-linked accounts before funds are lost. The initiative, announced on May 4, 2026, marks the first time MAS has formally pooled bank transaction data at the regulator level for pre-emptive financial crime detection.
Singapore’s financial system processes tens of millions of retail transactions daily across a dense network of digital payment rails, including PayNow, FAST, and card-based infrastructure. That volume creates both opportunity and exposure: scam syndicates exploit the speed of digital transfers to move funds out of reach within minutes of a victim authorising a payment. MAS’s Proof-of-Value exercise is designed to test whether AI models trained on pooled, multi-bank data can detect the signatures of scam activity before those transfers complete.
Structure Of The Proof-Of-Value
MAS announced on May 4, 2026 that it is conducting a Proof-of-Value exercise, working alongside the Government Technology Agency of Singapore and the Singapore Police Force, with five participating banks contributing historical transaction data. The identity of the five banks has not been disclosed in the MAS announcement. The POV draws on transaction records including bank account numbers to train and evaluate AI and machine learning models for their capacity to flag higher-risk transactions and accounts ahead of scam losses occurring.
The rationale for pooling data across institutions is statistical. Individual banks hold transaction histories for their own customers only, which limits the pattern recognition potential of any models trained on a single institution’s records. By combining datasets from five banks, MAS aims to give the models broader visibility into cross-institution transaction flows, which is where scam networks frequently route funds to obscure their origin.
Data Governance And Cryptographic Safeguards
MAS has established a secure data-sharing environment specifically for the exercise, governed by protocols designed to protect customer information while enabling collaborative model development. Bank account numbers included in the datasets are hashed using a one-way algorithmic process, converting them into unique encoded values that prevent direct identification. Only the originating bank retains the ability to identify the underlying account from its own hash. Access to the data environment is restricted to authorised personnel operating within a controlled and continuously monitored setting. All data used in the exercise will be deleted upon conclusion of the POV.
The cryptographic safeguards reflect Singapore’s established approach to regulated data sharing in the financial sector, where MAS has consistently required institutions to balance analytical utility against customer privacy obligations.
Scope And Potential Expansion
The current POV is framed by MAS as groundwork for deeper industry collaboration rather than a standalone pilot. If the exercise demonstrates that cross-institutional AI models produce measurable improvements in scam detection accuracy or speed, MAS has indicated it may expand the scope of the initiative to incorporate broader datasets and a wider set of use cases. The regulator stated explicitly that the POV is designed to complement, not replace, the existing anti-financial crime systems operated individually by each participating bank.
The initiative sits within a broader MAS push to apply AI to systemic financial crime challenges. MAS previously concluded phase two of Project MindForge, an industry consortium initiative that produced an AI Risk Management Toolkit for the financial services sector, with participants including DBS Bank, OCBC Bank, United Overseas Bank, Standard Chartered Bank, Citi Singapore, and HSBC.
Why Pre-Emptive Detection Matters
The distinction between reactive and pre-emptive scam detection is operationally significant. Existing fraud detection systems at Singapore banks are largely designed to flag suspicious transactions at the point of execution, triggering alerts that may or may not result in a hold being placed on the transfer before funds leave the institution. Once a payment clears into a mule account, recovery rates drop substantially.
Pre-emptive detection, by contrast, aims to identify accounts or transaction patterns that carry elevated scam risk before a victim initiates a transfer. That could mean flagging receiving accounts that exhibit characteristics consistent with mule account profiles, or identifying sequences of incoming transfers that match known scam fund-flow patterns. If the AI models developed through the POV prove accurate enough for operational deployment, they could be integrated into real-time payment screening infrastructure across Singapore’s banking sector.
The Singapore Police Force’s involvement in the exercise is notable. Law enforcement data on confirmed scam cases, including account-level information on convicted mule accounts, represents a ground-truth dataset that can be used to train and validate AI models in ways that bank transaction data alone cannot support.
Publicly available information on the specific transaction volumes, scam typologies, or model architectures being tested in the POV remains limited at this stage, consistent with MAS’s practice of releasing technical detail only after conclusion of pilot exercises.